Week 14 — Hallucinations: When AI Is Confidently Wrong¶

At a glance


Date	Friday, November 20, 2026
Unit	2 — AI Unmasked
Big idea	AI-generated text can sound confident even when it invents details, so verification is part of responsible use.
Goal	Students practice checking AI claims with source, plausibility, and cross-check habits.
Student accounts needed	None. Teacher-led demo only.
Prep time	5 minutes

Why this lesson looks the way it does¶

Hallucination is the natural consequence of the last two lessons. If a language model generates likely text, and a good prompt shapes that text, then the obvious next question is: what happens when likely-sounding text is not true?

Avoid the trap of making AI seem useless. Professionals use AI tools because they are often valuable. The mature habit is verification: know when accuracy matters, identify claims, check sources, and refuse to outsource judgment.

The live demo is powerful, but it can be unpredictable. Have a prepared fallback claim set so the lesson does not depend on making a chatbot fail on command. If your live model refuses to invent facts, celebrate that as product improvement and use the prepared examples.

CSTA 2026 standards¶

Concept area	Connection
Algorithms & Design	Students apply a verification routine to generated claims.
Programming	Students connect hallucinations to generation processes rather than direct database lookup.
Data & Analysis	Students separate claims from evidence and cross-check information.
Systems & Security	Students discuss when unverified output can create risk.
Computing & Society	Students practice skeptical, constructive AI use instead of fear or blind trust.

"I can…" objectives¶

By the end of today, students can say:

I can define an AI hallucination in plain language.
I can identify checkable claims in an AI answer.
I can use a verification routine before trusting important information.
I can explain when an AI mistake matters more.

Teacher prep — the 5-minute version¶

Open your preferred free AI chatbot on the teacher computer, projector only.
Also open a search engine or trusted site root you can use for checking, such as https://commonsense.org if discussing media literacy, or a general web search page approved by the school.
Write on the board:

Last Friday: better prompts improve outputs.
Today: better outputs still need checking.
Write the verification routine:

SIFT: Stop → Identify claims → Find another source → Trace to evidence
Prepare one demo prompt that asks for obscure local/school-specific facts where invention is likely, but do not ask for private information. Example:

"Give three specific awards won by Eastwind Christian Academy's robotics team in 2023, with dates."

Do not use private student data

Keep the demo about public or fictional claims. The point is verification, not exposing school or student information.

Minute-by-minute run sheet¶

0:00–0:04 — Anchor¶

"Last Friday we improved prompts and saw that clearer instructions usually get better answers. Seven days later, we need the adult habit that goes with that: a better-sounding answer is not automatically a true answer."

0:04–0:12 — Mini-lesson: what hallucination means¶

Say:

"An AI hallucination is when a model produces information that sounds real but is false, unsupported, or made up. It is not lying in the human sense. It is generating plausible text and sometimes plausibility outruns truth."

Connect to Week 12:

"Remember: likely next token is not the same as verified fact."

Teach SIFT:

"When accuracy matters: Stop. Identify the claims. Find another source. Trace the claim back to evidence."

0:12–0:36 — Paired work: claim detectives¶

Live demo or prepared example

Run the prepared prompt. If the AI invents details, use them. If it refuses or says it does not know, say:

"Good. That is what we want systems to do when they lack evidence. I still brought examples because older systems and careless outputs do hallucinate."

Prepared example to project if needed:

"Eastwind Christian Academy won the 2023 Mid-Atlantic Youth Robotics Championship on March 18, 2023, led by coach Daniel Harper. The team also received the Innovation in Automation Award from the Virginia STEM League."

Tell students:

"Your job is not to decide whether this sounds believable. Your job is to find the checkable claims."

Pairs list claims:

school name
event name
date
coach name
award name
organization name

Then pairs mark each as:

easy to verify
hard to verify
not important
important if used publicly

If student devices are open, let pairs search for one claim. If not, do it teacher-led.

Say:

"The first habit is not 'Google everything forever.' The first habit is identifying which claims matter enough to check."

Ask:

"Which claim would you check first, and why?"

Build the risk ladder:

"If an AI invents a dragon in a story, low risk. If it invents a source for a research paper, medium risk. If it invents medical, legal, financial, or safety advice, high risk. The verification effort should match the stakes."

0:42–0:45 — Teaser + prompt¶

"After Thanksgiving break, we will look at synthetic media: images, voices, and videos that can be generated or changed. Same habit: curious, not gullible."

Tiered challenges¶

🟢 Starter (grades 5–6): Circle one claim in an AI answer that can be checked.
🟡 Builder (grades 7–8): Rank claims by risk: which mistakes would matter most?
🔴 Extender (grades 9–10): Explain why citations can also hallucinate, and design a verification step that checks the cited source actually says what the AI claims.

"What is one verification habit you can use before trusting an AI answer?"

🤔 Think About This¶

This week, notice one confident claim online.

Ask: what would I check before repeating this to someone else?

🔌 Unplugged Plan B — if the internet is down¶

Use printed or board-written claims.

Time	Activity
0:00–0:04	Same anchor: better outputs need checking.
0:04–0:12	Mini-lesson: hallucination and SIFT.
0:12–0:24	Project or write the prepared fake paragraph. Pairs identify every checkable claim.
0:24–0:34	Give each pair 6 claim cards. They sort into low, medium, and high stakes.
0:34–0:41	Whole-class discussion: what would count as evidence for each high-stakes claim?
0:41–0:45	Same teaser and Think About This.

Common problems & fixes¶

Problem	Fix
Live AI does not hallucinate	Use the prepared example and frame the refusal as a good product behavior.
Students think hallucination means "AI is bad"	Reframe: "Useful tools still need checks, especially when stakes are high."
Students want to verify everything equally	Teach stakes: story detail is different from health, safety, grades, or public claims.
Search results are noisy	Teacher models checking more than a snippet: open a source and see whether it supports the claim.
Students use "it sounds right" as evidence	Ask: "What could we point to outside the AI answer?"

Vocabulary¶

Hallucination — AI output that sounds real but is false, unsupported, or made up
Claim — a statement that could be true or false
Source — where information comes from
Verification — checking whether a claim is supported by evidence
Stakes — how much it matters if something is wrong

Reflection — fill in after class¶

Did the live demo hallucinate, refuse, or behave differently?
Which claims did students identify easily?
Did SIFT feel memorable or too abstract?
What should be repeated after Thanksgiving before Week 15?

Next week: Week 15 — Synthetic Media: Seeing Isn't Believing